Agent Overview
Agent Composition
Section titled “Agent Composition”This project consists of 3 independent AI agents. Each agent is built on the Strands SDK and runs on AWS Bedrock AgentCore.
| Agent | Role | Model | Interface |
|---|---|---|---|
| IDP Agent | Document analysis, search, artifact generation | Claude Opus 4.6 | HTTP streaming |
| Voice Agent | Real-time bidirectional voice conversation | Nova Sonic | WebSocket |
| Web Crawler Agent | Web page crawling and content extraction | Claude Sonnet 4.6 | SQS trigger |
Common Architecture
Section titled “Common Architecture”User request │ ▼AWS Bedrock AgentCore │ ├─ AgentCore Runtime (ECS container) │ └─ Strands Agent │ ├─ LLM (Bedrock) │ ├─ MCP tools (AgentCore Gateway) │ └─ Built-in tools (Strands SDK) │ ├─ MCP Gateway (IAM SigV4 auth) │ ├─ Search MCP (LanceDB hybrid search) │ ├─ Graph MCP (Neptune graph traversal) │ ├─ Image MCP (image analysis) │ ├─ QA MCP (QA analysis management) │ └─ Document MCP (PDF/DOCX/PPTX/MD) │ └─ Code Interpreter (Python sandbox)MCP (Model Context Protocol)
Section titled “MCP (Model Context Protocol)”Agents access external tools via MCP. The AgentCore Gateway hosts MCP servers, and each agent calls tools through SigV4-authenticated HTTP connections.
All MCP tool calls automatically inject user_id and project_id, ensuring data isolation between users.
Session Management
Section titled “Session Management”IDP Agent and Voice Agent persist conversation history to S3.
s3://session-storage-bucket/└── sessions/ └── {user_id}/ └── {project_id}/ └── {session_id}/Agent Details
Section titled “Agent Details”- IDP Agent — Document search, analysis, artifact generation (DOCX/XLSX/PPTX/charts/diagrams)
- Voice Agent — Real-time voice conversation with Nova Sonic
- Web Crawler Agent — Web crawling with AgentCore Browser, D2Snap HTML compression